Robotic manipulators are reaching a state where we could see them in household environments in the following decade. Nevertheless, such robots need to be easy to instruct by lay people. This is why kinesthetic teaching has become very popular in recent years, in which the robot is taught a motion that is encoded as a parametric function - usually a Movement Primitive (MP)-. This approach produces trajectories that are usually suboptimal, and the robot needs to be able to improve them through trial-and-error. Such optimization is often done with Policy Search (PS) reinforcement learning, using a given reward function. PS algorithms can be classified as model-free, where neither the environment nor the reward function are modelled, or model-based, which can use a surrogate model of the reward function and/or a model for the dynamics of the task. However, MPs can become very high-dimensional in terms of parameters, which constitute the search space, so their optimization often requires too many samples. In this paper, we assume we have a robot motion task characterized with an MP of which we cannot model the dynamics. We build a surrogate model for the reward function, that maps an MP parameter latent space (obtained through a Mutual-information-weighted Gaussian Process Latent Variable Model) into a reward. While we do not model the task dynamics, using mutual information to shrink the task space makes it more consistent with the reward and so the policy improvement is faster in terms of sample efficiency.

In the next few years, the amount and variety of context-aware robotic manipulator applications is expected to increase significantly, especially in household environments. In such spaces, thanks to programming by demonstration, non-expert people will be able to teach robots how to perform specific tasks, for which the adaptation to the environment is imperative, for the sake of effectiveness and users safety. These robot motion learning procedures allow the encoding of such tasks by means of parameterized trajectory generators, usually a Movement Primitive (MP) conditioned on contextual variables. However, naively sampled solutions from these MPs are generally suboptimal/inefficient, according to a given reward function. Hence, Policy Search (PS) algorithms leverage the information of the experienced rewards to improve the robot performance over executions, even for new context configurations. Given the complexity of the aforementioned tasks, PS methods face the challenge of exploring in high-dimensional parameter search spaces. In this work, a solution combining Bayesian Optimization, a data-efficient PS algorithm, with covariate Gaussian Process Latent Variable Models, a recent Dimensionality Reduction technique, is presented. It enables reducing dimensionality and exploiting prior demonstrations to converge in few iterations, while also being compliant with context requirements. Thus, contextual variables are considered in the latent search space, from which a surrogate model for the reward function is built. Then, samples are generated in a low-dimensional latent space, and mapped to a context-dependent trajectory. This allows us to drastically reduce the search space with the covariate GPLVM, e.g. from 105 to 2 parameters, plus a few contextual features. Experimentation in two different scenarios proves the data-efficiency and the power of dimensionality reduction of our approach.